Information Technology Reference
In-Depth Information
Buffered
thumbnail video
frames; at least
up to frame
n+d
RoI trajectory
history up to
frame n
Motion analysis
Trajectory predictor
Predicted
RoI center at
frame n+d
Fig. 4
Video-content-aware RoI prediction analyzes motion in the buffered thumbnail video
frames. The video transmission system ensures that some thumbnail video frames are sent
ahead of time. Although not shown in the figure, RoI prediction can alternatively be per-
formed at the server. In this case, the server can analyze motion in the high-resolution frames,
however, the available trajectory history might be older than current due to network delays.
Also, the load on the server increases with the number of clients.
feature tracker [63]. The work in [66] is related in spirit, although the context,
mobile augmented reality, is different. In this work, MVs are used to track mul-
tiple feature points from one frame to the next while employing homography
testing to eliminate outliers among tracked feature points. The algorithm also
considers the case of B frames.
3. One can employ multiple RoI predictors and combine their results, for example,
through a median operation. This choice guarantees that for any frame-interval,
if one of the predictors performs poorly compared to the rest, then the median
operation does not choose that predictor. In general, the more diversity among
the predictors the better.
Compared to the video-content-agnostic schemes, the gain obtained through video-
content-aware RoI prediction is higher for longer prediction lookahead
d
[63].
Moreover, unlike the above approaches that are generic, the motion analysis can
be domain-specific [64]. For example, for interactive viewing of soccer, certain
objects-of-interest like the ball, the players, the referees, etc. can be tracked and
their positions can drive the RoI prediction.
In the approaches above, the user actively controls the input device and the goal
of the system is to predict the future path as accurately as possible. In another mode
of operation, the system offers to track a certain object-of-interest for the user such
that it relieves navigation burden. In this case, a user-selected trajectory might not
be available for comparison or as trajectory history input. In this mode, the goal of
the algorithm is to provide a smooth trajectory without deviating from the object.
Figure 5 reproduces a result from [64] that shows the tracking of a soccer
player over successive frames of the thumbnail video. The algorithm is based on
